RoRA: Efficient Fine-Tuning of LLM with Reliability Optimization for Rank Adaptation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Liu, Jun, Kong, Zhenglun, Dong, Peiyan, Yang, Changdi, Shen, Xuan, Zhao, Pu, Tang, Hao, Yuan, Geng, Niu, Wei, Zhang, Wenbin, Lin, Xue, Huang, Dong, Wang, Yanzhi
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912183999266816
author Liu, Jun
Kong, Zhenglun
Dong, Peiyan
Yang, Changdi
Shen, Xuan
Zhao, Pu
Tang, Hao
Yuan, Geng
Niu, Wei
Zhang, Wenbin
Lin, Xue
Huang, Dong
Wang, Yanzhi
author_facet Liu, Jun
Kong, Zhenglun
Dong, Peiyan
Yang, Changdi
Shen, Xuan
Zhao, Pu
Tang, Hao
Yuan, Geng
Niu, Wei
Zhang, Wenbin
Lin, Xue
Huang, Dong
Wang, Yanzhi
contents Fine-tuning helps large language models (LLM) recover degraded information and enhance task performance. Although Low-Rank Adaptation (LoRA) is widely used and effective for fine-tuning, we have observed that its scaling factor can limit or even reduce performance as the rank size increases. To address this issue, we propose RoRA (Rank-adaptive Reliability Optimization), a simple yet effective method for optimizing LoRA's scaling factor. By replacing $α/r$ with $α/\sqrt{r}$, RoRA ensures improved performance as rank size increases. Moreover, RoRA enhances low-rank adaptation in fine-tuning uncompressed models and excels in the more challenging task of accuracy recovery when fine-tuning pruned models. Extensive experiments demonstrate the effectiveness of RoRA in fine-tuning both uncompressed and pruned models. RoRA surpasses the state-of-the-art (SOTA) in average accuracy and robustness on LLaMA-7B/13B, LLaMA2-7B, and LLaMA3-8B, specifically outperforming LoRA and DoRA by 6.5% and 2.9% on LLaMA-7B, respectively. In pruned model fine-tuning, RoRA shows significant advantages; for SHEARED-LLAMA-1.3, a LLaMA-7B with 81.4% pruning, RoRA achieves 5.7% higher average accuracy than LoRA and 3.9% higher than DoRA.
format Preprint
id arxiv_https___arxiv_org_abs_2501_04315
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle RoRA: Efficient Fine-Tuning of LLM with Reliability Optimization for Rank Adaptation
Liu, Jun
Kong, Zhenglun
Dong, Peiyan
Yang, Changdi
Shen, Xuan
Zhao, Pu
Tang, Hao
Yuan, Geng
Niu, Wei
Zhang, Wenbin
Lin, Xue
Huang, Dong
Wang, Yanzhi
Machine Learning
Artificial Intelligence
Fine-tuning helps large language models (LLM) recover degraded information and enhance task performance. Although Low-Rank Adaptation (LoRA) is widely used and effective for fine-tuning, we have observed that its scaling factor can limit or even reduce performance as the rank size increases. To address this issue, we propose RoRA (Rank-adaptive Reliability Optimization), a simple yet effective method for optimizing LoRA's scaling factor. By replacing $α/r$ with $α/\sqrt{r}$, RoRA ensures improved performance as rank size increases. Moreover, RoRA enhances low-rank adaptation in fine-tuning uncompressed models and excels in the more challenging task of accuracy recovery when fine-tuning pruned models. Extensive experiments demonstrate the effectiveness of RoRA in fine-tuning both uncompressed and pruned models. RoRA surpasses the state-of-the-art (SOTA) in average accuracy and robustness on LLaMA-7B/13B, LLaMA2-7B, and LLaMA3-8B, specifically outperforming LoRA and DoRA by 6.5% and 2.9% on LLaMA-7B, respectively. In pruned model fine-tuning, RoRA shows significant advantages; for SHEARED-LLAMA-1.3, a LLaMA-7B with 81.4% pruning, RoRA achieves 5.7% higher average accuracy than LoRA and 3.9% higher than DoRA.
title RoRA: Efficient Fine-Tuning of LLM with Reliability Optimization for Rank Adaptation
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2501.04315